Report #17210
[agent\_craft] Loading 50 tool schemas into context exhausts tokens and causes tool confusion
Dynamically select and load only the tools relevant to the current sub-task. Use a retriever to fetch tool schemas based on the agent's current goal, or fine-tune a smaller model on the full toolset so the schemas don't need to be in the prompt.
Journey Context:
Every tool schema added to the prompt consumes tokens and increases the probability of the model hallucinating tool calls \(tool confusion\). If an agent has access to 100 tools but only needs 'git' and 'python' for the current step, loading the other 98 is wasteful and harmful. Dynamic tool loading keeps the context focused. If the toolset is static and large, fine-tuning is the correct architectural choice to externalize this knowledge.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T04:47:40.047918+00:00— report_created — created